Research papers
C) For peer reviewed paper (submitted)
3. Data material
This study uses a range of multispectral and multiresolution remote sensing images (Landsat-8, WorldView-3) combined with spatial vector data (GIS building footprint). To derive the surface temperature (°C) of the study area, a cloud free Landsat 8 scene (path 162, row 38) was provided free of charge on the U.S. geological survey (USGS) website (www.earthexplorer.usgs.gov). The spatial resolution of Landsat 8 is 30 m for the operational land manager (OLI) bands. A spatial resolution of 100 m for the thermal infrared sensor (TIRS) was resampled to 30 m by the USGS to match the OLI bands (Table 1). In addition, Worldview-3 cloud free satellite images with 1.5 m multispectral spatial resolution were used to analyze the land cover of the study area. This geospatial data were prepared from high precision aerial images (less than 1 m resolution), including the height of each building
(Municipality of Yazd, 2016). The building layer was used for three-dimensional (3D) modeling of the entire study area.
Table 1 Details of remote sensing image and spatial vector data.
Sensor Resolution (m) Acquisition Date
Landsat 8 30 m (OLI) & 100 m (TIRS) 9 August 2016 Worldview-3 1.24 m (multispectral resolution) 16 August 2016
Building footprint vector 1 m 1 June 2016
4. Methods
To study the effect of urban 3D morphology on surface temperature, multiple 2D and 3D indicators were incorporated in a regular spatial grid (section 4.4). A 30-m grid size cell was considered to properly reflect multiple indicators with the resolution of surface temperature pixels over space. A machine learning algorithm was used to identify the association between the dependent variable (surface temperature) and independent variables (section 4.5). The following flowchart (Figure 2) illustrates the methodology.
4.1 Surface temperature
From a remote sensing point of view, the ‘surface’ is whatever the ground sees when it looks through the atmosphere. The surface could be a grassland, forest canopy or the roof of a building. Moreover, the surface temperature is the radiative skin temperature of the land surface that is measure in the direction of the airborne sensor (satellite). Surface temperature represents the average surface temperature in a given unit area at a certain time of day, instead of measuring the temperature of the air, like weather stations do. The surface temperature depends on the Albedo, the vegetation cover and the soil moisture.
To derive the surface temperature, the Landsat 8 TIRS bands (bands 10: 10.60 µm to 11.19 µ and band 11: 11.50 µm to 12.51 µm) were used to convert the digital numbers to space reaching radiance on top of the atmosphere measured by the instrument. The two bands in the long-wave infrared (LWIR) use Quantum Well Infrared Photodetector technology to detect long wavelengths of light emitted by the Earth surface at 100-meter resolution. A description of a step-by-step process proposed by the USGS (Landsat 8, Handbook, 2015) was used to create surface temperature maps.
The retrieval of surface temperature from satellite thermal bands can be broadly classified into three categories: single-channel method (mono-window algorithm), multi-channel methods (split-window algorithm) and multi-angle methods. The methodology suggested by Yuan and Bauer (2007) and Du et al. (2015) were used to estimate the surface temperature from Landsat 8 TIRS data by applying a split-windows algorithm. The split-window algorithm removes the atmospheric effect in accordance with the different absorptions of two TIRS bands, and the linear or nonlinear combination of the brightness temperatures is finally applied for the surface temperature.
4.2 Land cover classes
For mapping land cover classes in the city of Yazd, a supervised learning algorithm called support vector machines (SVM) was used. SVM are supervised non-parametric statistical learning techniques that improve classification and regression problems. Methods based on SVM have been particularly successful in applications with a large set of variables (Tax and Duin, 2004), such as remote sensing data. The methodology proposed by Okujeni et al., 2013 was implemented to classify the Worldview-3 image of the study area into vegetation, impervious surface, bare soil and water. First, three classes were used as the three main land
cover types in the study area. The last class (water) was defined to ensure that the surface temperature of the study area was not affected by the water bodies. For validation of results, an accuracy assessment was conducted by measuring the overall accuracy with 255 randomly selected control points.
4.3 Three-dimensional city model
A 3D city model was required to determine the characteristics and geometry of the built-up environment. GIS geospatial data (building footprint) were used to shape the 3D form of the city. To do this, each individual building height was combined with the building footprint. Buildings are delineated as box-shaped 3D form. This is an accepTable approximation of the real building shape, considering the flat roofs in Yazd. To compute the 3D model of the urban environment of Yazd, ArcGIS 10.3 and ArcScene 10.3 were used. To study the urban 3D environment, the following 3D indicators were considered: the volume of each individual building (m3), the surface area of each individual building (m2), the
building rooftop area (m2), the shadow footprint (m2) of each building based on the height
of the building at the time of satellite image acquisition, the shadow volume (m3) of each
building based on the height of building at the time of satellite image acquisition and the ratio of building’s height to its width (H/W) (Figure 3).
The volume of each building in cubic meters (m3) and the surface area of each building in
square meters (m2) are calculated using the following equations:
𝑉𝑉𝑉𝑉𝑉𝑉𝑉𝑉𝑉𝑉𝑉𝑉 = 𝑊𝑊 × 𝐷𝐷 × 𝐻𝐻 equation 1)
𝐵𝐵𝑉𝑉𝐵𝐵𝑉𝑉𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵 𝑠𝑠𝑉𝑉𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑉𝑉 𝑠𝑠𝑠𝑠𝑉𝑉𝑠𝑠 = 2(𝐻𝐻 × 𝑊𝑊) + 2(𝐻𝐻 × 𝐷𝐷) + (𝑊𝑊 × 𝐷𝐷) equation 2) where W is the width of the building, D is the depth of the building and H is the height of the building. Then, the surface area of each individual building, i.e., the sum of the surface areas of all of the walls, including the rooftop area, was determined (Figure 3).
Figure 3 The 2D shadow footprint and the 3D shadow volume.
4.4 Data integration and statistical modeling
The data sources used in this research have different spatial resolution. Therefore, an appropriate way to standardize data is to superimpose a grid with a common spatial resolution of 30 m × 30 m. This grid size is equal to the surface temperature pixel size and has proved to be an optimal size for aggregating 2D and 3D urban feature variables. A regular spatial grid was used to aggregate all of the 2D and 3D information into a homogeneous database. Table 2 shows the list of 2D and 3D indicators and their distribution ranges that were superimposed with the regular grid. The homogenous database offers a convenient structure for the subsequent boosted regression tree analysis. Due to the atmospheric flows and exchanges, the observed temperature at a given grid-cell is likely to be spatially correlated to the neighboring cells. Figure 4 shows the integration of the multi-dimensional indicators with the regular spatial grid.
4.5 Statistical model
A machine learning statistical model called a boosted regression tree (BRT), also known as boosted decision tree, was used to analyze the association of multi-dimensional indicators (2D and 3D) in the complex urban environment. The BRT approach differs fundamentally from conventional techniques that use a tool for quantifying the relationship between one variable and others upon which it depends. This models is advanced for analysis of the morphological relationships (Clarke & Johnston, 1999). The BRT combines the strengths of
two algorithms, regression tree and boosting, in a single performance (Friedman, 2002). Regression trees are from the classification and regression trees (decision trees) group of models, and boosting builds and combines a collection of models. Boosting means that each tree is dependent on prior trees, and learns by fitting the residual of the trees, and the decisions respectively, that preceded it. The BRT method is a supervised learning method which requires a labeled dataset containing numerical values. The model can be trained by providing the model and the labeled as an input into the model. Thereafter, the trained model could be used to predict values for the new input examples. BRT could handle different types of predictor variables and accommodating missing data.
Figure 4 Integrating the multi-dimensional indicators into a regular spatial grid with a common spatial resolution of 30 m × 30 m.
More specifically, the relative influence measure was scaled from 0 to 100%. The scale relies on the number of times an independent variable is selected for splitting during the decision tree process and weighted by the squared improvement to the model as a result of each split (Friedmann and Meulman, 2003). For our study, 50% of the data were used to fit our dependent variable and the surface temperature data as the first regression tree (bagging factor of 0.5). We included a “Laplace” distribution to fit our model because this kind of probability distribution can handle the non-binary surface temperature with multiple temperature values
available (N/A). For instance, not all of the grid cells had vegetation cover (2D); therefore, they were designated N/A. The regression analysis was conducted on both districts separately to compare the association of their urban structure and geometry with the surface temperature. The results were evaluated using the scaled measures of the relative influence calculation. The statistical programming software R version 3.3.1 (2016-06-21) "Bug in your hair" with the main packages of "raster," "mmand," "rgdal", "gbm", and "dismo" (RCoreTeam 2015) was used to conduct the analysis on 24.978 Million entries. A notebook with Intel Core i5-6300U processor of 6th generation (2.4 GHz, 3 MB Cache) and 16 GB of DDR4 RAM (2.133 MHz) were used to process the R script in 6 hours.
Table 2 Descriptive statistics for urban feature variables on the grid structure for historic and new structures of the study area.
Urban feature variable Indicator Unit
New district Historic district Min. value Max. value Min. value Max. value Building height 3D m 0.0 36.0 0.0 5.9 Building volume 3D m3 0.0 13004.4 0.0 6523.18
Building surface area 3D m2 0.0 3324.1 0.0 3603.45
Shadow footprint 3D m2 0 7679.85 0 2174.54
Shadow volume 3D m3 0.0 1843.0 0.0 342.1
H/W --- n/a 0.0 11.6 0.0 1.7
Surface temperature 2D °C 29.3 50.4 34.4 45.1
Building rooftop area 2D m2 0.0 899.3 0.0 648.8
Street 2D m2 0.0 890.6 0.0 551.1
Vegetation 2D m2 0.0 900.0 0.0 850.0
5. Results
5.1 Land cover analysis and surface temperature
Cities have a complex environment that includes buildings, roads, and green and blue structures. Figure 5 illustrates the result for SVM classification with an overall accuracy assessment of 86.5%. Generated land cover data were overlaid with Landsat 8 thermal band data to retrieve the surface temperature of the study area.
Figure 5 SVM land cover classification with an overall accuracy assessment of 86.5%. The left image of the upper box shows the Worldview-3 image of the new structure, and the right-hand side shows the SVM classified image. The left image of the lower box shows the high resolution satellite image of the historic district, and the right-hand side shows the SVM classified image of the same area.
Figure 6 (A) shows the daytime surface temperature of the city of Yazd and its surrounding on 9 August 2016 at 10.30 A.M. local time. The number of blue boxes in Figure 6 (A) shows the association of different land cover types with the surface temperature. The land cover type
of Yazd is cooler than the surrounding suburb. This result contradicts the current definition of UHI that argues that cities have warmer temperatures than the surrounding areas due to human activity. Figure 6 (B) shows that the thermal band of Landsat 8 distinguishes the historic district from the new district. Figure 6 (A,B) shows a warmer surface temperature in the historic district than in the newly built district. The highest surface temperature recorded in the historic district was 45 °C, whereas that in the new district was 39 °C.
Figure 6 Surface temperature of Yazd on 9 August 2016 at 10.30 A.M. local time. A) The inner urban area shows a lower surface temperature in contrast to its suburban area. The land cover of the numbered blue boxes is as follows: 1, historic district; 2, central rail way and open field; 3, bare field; 4, 5, and 6, newly structured district. B) Infrastructure in the historic district is more affected by heat. The average surface temperatures of the historic and new districts are 40 °C and 37 °C, respectively.
5.2 3D city model
The 3D spatial models of both the new and historic districts of Yazd are shown in Figure 7. The well-aligned structure of the new district has abundant urban vegetation, and the compact and dense structure of the historic district has bare vegetation cover.
Figure 7 Yazd 3D model including the shadowing effect of each individual building. A) 3D model of the new district; B) 3D model of the historic district.
5.3 Boosted regression tree analysis
The BRT was processed independently for both districts to understand the relative spatial effect of urban multi-dimensional indicators (2D and 3D) on surface temperature. Figure 8 illustrates the BRT analysis of the 2D and 3D indicators for both districts in separate graphs. The left-hand graph shows the percentage of the relative influence of the 2D and 3D indicators on the surface temperature for the historic district (HD). The right-hand graph illustrates the same information for the new district (ND). HD1-5 depicts the percentage of different 3D
indicators for the historic district. HD6-9 shows the percentage for different 2D indicators of
the historic district. ND1-5 illustrates the percentage of different 3D indicators in the new
district, and HD6-9 illustrates the percentage of the different 2D indicators of the new district.
The results show that all of the 2D and 3D indicators play a role in defining the surface temperature. However, some indicators have a greater effect than others. The overall relative influence on surface temperature in the historic district (HD6-9) was 30% and that for 3D
indicators in the same district (HD1-5) was 70%. The overall relative influence on surface
temperature for the new district were 23.5% for the 2D indicators (ND6-9) and 76.5% for the
3D indicators (ND1-5). In other words, 3D indicators have a much stronger relative influence
The relative influences on surface temperature for 3D indicators of the historic district are as follows. Building height has the strongest relative influence (HD1), with 21%. This is followed
by the building H/W ratio (HD2), with 14%; building shadowing effect (HD3), with 13%;
building surface area (HD4), with 12%; and building volume (HD5), with 10% of relative
influence on the surface temperature. The relative influence of 2D indicators on surface temperature in the historic district is weaker. The strongest relative influence among 2D indicators on surface temperature is from the street area (HD6), with 14%. This is followed by
vegetation cover (HD7), with 9%; building rooftop area (HD8), with 4%; and land cover type
(HD9), with 3%.
Figure 8 The relative influence of urban multi-dimensional indicators on the surface temperature of the city of Yazd. The left-hand graph shows the analysis for the historic district (HD). HD1: building
height with 21%, HD2: building H/W ratio with 14%, HD3: building shadow with 13%, HD4: building
surface area with 12%, HD5: building volume with 10%, HD6: Building street area with 14%, HD7:
vegetation cover with 4%, HD8: building rooftop area with 4%, and HD9: urban land cover type with
3%. The right-hand graph illustrates it for the new district (ND). ND1: building surface area with 30%,
ND2: building height with 24.5%, ND3: building shadow with 18%, ND4: building H/W ratio with 2%,
ND5: building volume with 2%; ND6: building rooftop area with 18.5%, ND7: street area with 3%,
The relative influence of 3D indicators in the new district (ND1-5) also shows a stronger
association with surface temperature than the 2D indicators. With 30%, building surface area (ND1) is the strongest relative influential 3D indicator associated with surface temperature in
the new district. This is followed by building height (ND2), with 24.5%; building shadow
(ND3), with 18%; building H/W (ND4), with 2%; and building volume (ND5), with 2%. The
relative influence of 2D indicators on surface temperature in the new district (ND6-9) are as
follows: building rooftop area (ND6), with 18.5%; street area (ND7) and (ND8), with 3%; and
vegetated area (ND9), with 2%.
6. Discussion
3D studies across different fields have received great attention, but this technique is still in its infancy (Alavipanah et al., 2016). The focus of this research was to understand the variation in surface temperature caused by different urban multi-dimensional environment. A machine learning algorithm was used to understand the relative influence of 2D and 3D indicators on surface temperature.
Many previous UHI studies have confirmed that UHI is characterized by higher temperature in the urban area than in its surroundings. For example, the temporal and spatial characteristics of UHI were published by Cui et al. (2017); remote sensing of UHI across the continental USA was reported by Imhoff et al. (2010); and Deosthali (2000) studied the impact of rapid urban growth on UHI. However, our results show a cooler surface temperature for the inner urban area of Yazd compared to the surrounding suburbs. Another study in the semi-arid city of Erbil, Iraq, showed similar behavior (Rasul et al., 2015). This finding contradicts the current definition of UHI. The current definition is suiTable for cities in mid-latitude (Ward et al., 2016), as well as in tropical (Qaid et al., 2016) and sub-tropical (Lau et al., 2016; Goldreich, 1992) regions. The question that arises is whether the traditional and established definition of UHI is adequate to encompass all types of cities. Therefore, it is recommended to use the term ‘urban cooling island’ in studies when the heat pattern of the inner city is less than that of the surrounding suburb, particularly in arid and semi-arid cities.
sun light and can be linked to high surface temperatures. Second, the presence of vegetation in the urban area is responsible for several mechanisms of simultaneously cooling, similar to the evapotranspiration process and shading effect. Furthermore, regular irrigation of parks and gardens in the city, especially during the hot season, could reduce the temperature in the urban area. Shadowing effects are also additional factors. Based on solar position and the urban 3D geometry (buildings and vegetation), various angles of a building or vegetation receive light, and the other side remains comparatively dark. As the penetration of sunlight and energy storage in the soil is less for shaded areas, the temperature is also cooler. Therefore, shadowing could possibly lead to lower surface temperature within the cities.
In addition, the findings show that dissimilar urban structures, historic and newly built, lead to different spatial heat patterns. From the results, it is interesting to see that thermal data distinguished the historic district from the new district. In other words, the historic district and newly built district have different heat patterns. This difference may result from the different building materials and 3D spatial orientation of the buildings in both districts. The material used in the historic district is mainly adobe bricks (a mixture of clay soil, water and straw), whereas the material of the newly built district consists of cement and concrete. The thermal conductivity for adobe bricks (1.0 W/mK) is greater than the thermal conductivity of concrete (0.3 W/mK) at the room temperature (25 °C). It was probably due to this fact that ancient architects used two methods to avoid the penetration of sunlight into the buildings: building thick walls (from 1 to 1.5 meters) and mixing straw with mud when molding the bricks. Experience has shown that the existence of straw reduces solar penetration by reflecting solar radiation (Bahadori-nejad and Yaghoebi, 2006), so this could be a reason for the warmer heat pattern in the historic district.
The results of this research also depict the critical role that the 3D spatial orientation of the urban environment plays in shaping the heat pattern of the study area. The newly structured residential areas have a regular urban structure with aligned, multistoried buildings, wide streets and a relatively large amount of street vegetation. Wide streets and a regular urban structure facilitate better wind flow and air convection, as Wang and Akbari (2016) reported. In contrast, the 3D spatial configuration of the historic district is quite compact, with one-story buildings, narrow streets with continuous walls where the vegetation is limited to the central garden of each house. One should bear in mind that the high density of constructed buildings and streets was a strategy to avoid the penetration of hot solar radiation. The urban physical form, in particular the 3D building shape, plays a critical role in determining the amount of
incoming solar radiation (source and sink). Solar radiation increases the surface temperature of areas without any obstructions (building or vegetation cover). Nearby structures, including the Sahn, could benefit from building shadows during different times of the day, providing great relief from the hot sunlight and dry climate. However, perhaps the ancient architects were not aware of the phenomenon known as the urban street canyon, which intensifies the UHI effect by trapping the heat inside compact structures and blocking the wind. The focal